International Journal of Mineral Processing, 40 ( 1993 ) 1-16
1
Elsevier Science Publishers B.V., Amsterdam
Monitoring grinding parameters by signal measurements for an industrial ball mill Yigen Zeng and Eric Forssberg Dept. of Chemical and Metallurgical Engineering, Luled University of Technology, S-977 53, Lule& Sweden (Received 1 December 1992; accepted after revision 6 July 1993)
ABSTRACT Mechanical grinding emits a high-intensity vibration signal that contains information on the mill operating state. Vibration signals from the mill are presented in the form of both mechanical vibration and acoustic pressure. To apply these source signals to monitoring of grinding parameters, industrial scale grinding tests were performed with an iron ore at LKAB, Malmberget. Three operating parameters were considered: the feed rate, the mill feed size and the pulp density of mill discharge. The measured response parameters were the ground product size, the power draw and the pulp temperature. The source signals of the time-domain waveforms were simultaneously sensed by accelerometer and microphone so as to obtain a "stereograph" of grinding. The signals were first stored on a DAT recorder and then converted into digital format by an oscilloscope. The digitised waveforms were transformed into frequency-domain spectra by power spectral estimation. The variations on the power spectra can be described by a few "latent" variables derived by principal component analysis. Finally, close relations were established between key grinding parameters and "latent" variables by multiple regression. Using signal measurements, an automatic and efficient strategy can be developed to monitor operating parameters for the control system in a ball grinding circuit.
INTRODUCTION
It is common knowledge in the field of mineral processing that mechanical grinding produces strong acoustic noise and mechanical vibrations. The frequency pattern and the energies of acoustic/vibration signals contain information directly related to the operating state of grinding. The mill operator estimates the state of comminution based on the noise produced by the grinding mill. The accuracy of prediction on grinding comes through experience. To achieve a good grinding result or to come to a satisfactory decision, the operator's experience and association with the particular mill is of utmost importance. With the advent of the technology on signal acquisition, signal processing, system identification and advanced digital computation, the mill operator's decision-making can be imitated by computerised instrumentation. The source signals are simultaneously picked up by a vibrometer and an acoustic micro0301-7516/93/$06.00
© 1993 Elsevier Science Publishers B.V. All rights reserved.
2
Y. ZENG AND E. FORSSBERG
phone at key locations close to the mill. The original signals can then be amplified, stored and processed using digital technique. After some period of learning particulars of spectra (finger prints) from the grinding mill, the variation of the spectrum can be adopted to describe the changes of the operating state of the mill. The advantages of using a computerised instrument over human beings are obvious: the instrument can be set to work 24 hours a day on-line and may be situated in hazardous or tough working environments. After successful laboratory scale studies (Zeng and Forssberg, 1992, 1993 ), a series signal measurements were carried out on an industrial scale ball grinding. EXPERIMENTAL
Test material and grinding mill The material was fine crushed iron ore at LKAB mine in Malmberget, Sweden. The original ore was upgraded by primary magnetic separation before grinding. The particle size of the material fed to the ball mill was < l 0 mm. To apply the signal technique on monitoring parameters in grinding, industrial scale signal measurements were performed on a primary ball mill, ~ 4 . 2 X 5.4 m ( ~ 3 9 2 0 X 5200 m m inside the shell), with overflow discharge in an open circuit. Under normal operation, the feed rate to the magnetic separator was 255 t/h and the pulp density in the mill was 65% solids by weight. The mill speed was 12.8 rpm (corresponding to 60% of the critical speed) and the charge volume was about 35%. The grinding charge originally consisted of 100 tons of steel balls. The make-up balls of ~ 6 0 and ~ 4 0 m m in diameter, each accounting for 50% by weight, were continuously added to the mill to maintain the power draw level at 1130 kW.
Description of signal sensors and recorder A portable vibration meter, Model VM61, from Rion in Japan was equipped with the basic functions required to measure and evaluate a wide range of mechanical vibration. The vibration sensor was a shear structure piezoelectric type accelerometer with a built-in preamplifier. The basic signal, acceleration, can be transformed into the velocity and the displacement by requirement. The analogue signal can be output continuously through an AC terminal. Three lowpass ( 1, 5 and 15 kHz) and highpass ( 3, 10 and 1000 Hz) filters were built-in in the vibrometer with a cutoff attenuation at - 18 dB/ octave. To ensure the location and proper attachment of the vibration sensor, the accelerometer can be fastened on the measuring surface either by screws or by a magnet. Combination ofhighpass and lowpass filters was used to eliminate disturbances of the DC and high frequency resonances to the original
MONITORINGGRINDING PARAMETERSBY SIGNALMEASUREMENTS:INDUSTRIALBALLMILL
3
vibration signal. The sensitivity of original signal could be set to 0.1, 1 and 10 m / s 2.
A hypercardioid, condensing acoustic microphone, Model C1000S from AKG in Austria, was designed for professional vocal and instrumental use. The high-quality back plate condensing transducer works with the capsule shock mounted to optimise the handling and the cable noise rejection. The microphone case consisted of an aluminium shaft with a screw-on brass front tube and a rugged stainless steel mesh cap. The reliable frequency range was 50-20000 Hz with a sensitivity of 6 m V / P a ( - 4 4 . 4 dBV) at 1000 Hz. The output was an AC analogue signal. A digital audio tape recorder (DAT Model XD-Z505 B/E from JVC in Japan) with two channels was adopted for saving the original signal during tests. The unit was operated in a mode of sampling frequency at 48 kHz and 16 bits linear quantisation. The frequency response was 2-22000 Hz with the variations of ___0.5 dB. The signal-to-noise ratio was 92 dB. The analogue input terminal ranges from 63 to 500 mV at full scale. The signal format from the DAT can be output in either digital or analogue format.
Layout of grinding system and test procedure Fig. 1 shows the layout of drive system, ball mill and locations of signal sensors, signal amplifiers and DAT recorder. The fine crushed iron ore (0-10 m m ) was first fed to a wet magnetic separator before sending to the ball mill. When the magnetite content of the iron ore was stable, the real weight of iron ore fed to the ball mill was proportional to the weight transported by the belt conveyor. After magnetic separation, the water content in iron ore concen-
Ball mill
F
g
3
P
3
1. Drive motor 2. Gearbox 3. Pinion axis/gear 4. Trunnion bearings 5. Pinion bearings 6. Accelerometer 7. Vibration signal amplifier 8. Acoustic microphone 9. Acoustic signal amplifier 10. Digital audio tape deck Fig. 1. Layout of ball mill and signal acquisition instrument on dominate diagram.
4
Y. ZENG AND E. FORSSBERG
TABLE 1 Operating and response parameters in ball mill grinding Test No.
1 2 3 4 5 6 7 8
Feed rate (t/h)
148.4 147.3 145.8 151.4 235.2 234.8 235.6 236.1
Pulp density (%)
67.61 63.43 72.78 60.24 70.11 66.45 70.66 71.25
Power draw (kW)
1140 1127 1103 1113 1093 1108 1098 1073
Pulp temp. (°C)
20.6 19.9 21.8 19.3 17.8 17.0 18.0 17.6
Particle size distribution Feed (mm)
Product (ram)
Fso
Fso
Pso
Pso
3.44 5.61 7.35 6.74 1.58 4.24 4.16 8.48
0.710 0.808 0.823 0.841 0.586 0.715 0.602 1.161
0.220 0.210 0.222 0.212 0.333 0.309 0.313 0.347
0.099 0.097 0.098 0.102 0.167 0.155 0.151 0.166
trates was nearly constant. To adjust the pulp density, a certain amount of additional water was directly injected into the mill. The signal and material sampling commenced after an equilibrium period of 45 minutes if operating conditions were changed. Both vibration and acoustic signals were picked up, amplified and continuously recorded for 10 minutes. The material sampling follows the order: first the mill feed, then the pulp temperature and finally the mill discharge (the pulp density and the mill product size). The power draw and the feed rate were continuously recorded by the mill operating system. The particle size distribution of the mill feed and product were analysed by screening. The average values of grinding parameters are shown in Table 1.
Acquisition of acoustic and vibration signals The locations of signal sensors were investigated before formal signal measurements in industrial scale grinding. The location of the vibration sensor affects strongly the total signal energy, but only slightly the variation of the spectra. The location of the acoustic microphone was more sensitive than that of the accelerometer. The noise emission was different around the vertical cross section of the mill shell. To simplify the process, the mill shell can mainly be divided into abrasion and empty zones. When the microphone was located in the centre of the abrasion zone, the sensor location shifting along the mill axis affects strongly the total signal energy but only slightly the variation of spectra. Measuring variations of both mechanical vibration and acoustic pressure can provide a "stereograph" of grinding. To pick up sensitive signals with one accelerometer and one microphone, the accelerometer was vertically screwed on the bearing of a pinion axis and the microphone was placed at the
MONITORING GRINDING PARAMETERS BY SIGNAL MEASUREMENTS: INDUSTRIAL BALL MILL
5
centre of the abrasion zone, 200 mm away from the mill's outside shell. The original mechanical vibration and acoustic signals were amplified and stored in the L- and R-channel of the DAT recorder respectively. Due to limitations on hardware linkage, source signals must be digitised by a digital oscilloscope before being processed by an IBM compatible personal computer. To avoid aliasing by non-interesting high frequencies, the sampiing frequency of the A/D converter must be set at least 2.5 times the highest frequency of interest (Thomas, 1978). Primary studies show that the most important information can be obtained in the frequency range of 0-3 kHz; therefore the source signal should be passed through 4-cascade lowpass filter combination (each with a cutoff in - 2 4 dB/octave) at a cutoff frequency of 3 kHz before data acquisition. The sampling frequency of the oscilloscope was set to 10 kHz. Due to the oscilloscope's limitated memory, each signal sample corresponds to about 6.5 seconds absolute grinding period. The industrial scale grinding was operated under stable conditions; five signal samples were taken from the original full length signal recordings for each test.
Procedure of data analysis After data acquisition, time-domain waveforms of acoustic and vibration signals were transformed into frequency-domain spectra by Welch's method (Welch, 1967). This method involved sectioning the waveform data into either overlapping or non-overlapping sections. Each section (here 2048 points for one FFT section) was multiplied by an appropriate digital "window" function (here Hanning window) before the periodograms were computed. Finally, the modified periodograms were averaged and the resulting spectral estimate is asymptotically unbiased and consistent. To compensate voids on waveforms caused by the windowing function, the sections overlapped by 60% (DeFatta et al., 1988 ). The current problem on identifying a spectrum is that there are more frequency elements than measurements. About 600 frequencies were found at current frequency resolution. Since the energies of the spectral peaks (independent variables) were linearly correlated, principal component analysis (PCA) was suitable to reduce the independent variables into a few "latent" variables (Wold et al., 1987 ). By multiple regression, the relationship can be established between key grinding parameters and variation of the spectra. Digital signal processing and parameter identification were fulfilled by DSP4ME ® analytical computer software (developed by the first author) based on MatLab ® programming environment (Anon., 1990).
6
Y. ZENG AND E. FORSSBERG
RESULTS A N D D I S C U S S I O N
Partial correlation between grinding parameters The correlation coefficient is a measure of the degree of interrelation of two random variables. Partial correlation is an index of association of two variables holding constant or eliminating additional variables. It reflects a real correlation between two variables of interest. Table 1 shows the observed values of the operating parameters: mill feed rate (F, in t / h ) , feed size (mm) and pulp density (C, in % ), and response parameters: power draw (P, in kW), ground product size ( m m ) and pulp temperature (t, in ° C). Partial correlation between these parameters is analysed with 40 measurements and shown in Table 2. The correlation coefficient between pulp density and pulp temperature is 0.98, therefore these two parameters can be represented by each other. Feed rate is closely related to pulp density. No single parameter is significantly related to power draw. Therefore, power draw, feed rate, pulp density and mill product size were regarded as key grinding parameters.
Power spectra for vibration and acoustic signals The peak on the power spectrum depends on how it lines up with the bins. The area within the peak is therefore useful and represents the energy of signals (Little and Shure, 1988 ). Under identical frequency resolution, each peak on the power spectral plot implies a narrow frequency band. The frequencydomain power spectra for the vibration and acoustic signals are shown in Fig. 2. The total energies under all peaks of the spectra are similar for vibration and acoustic signals. The signal sensors were situated close to the mill shell; therefore signal peaks at about 115, 170, 227, 280, 337 and 450 Hz on the frequency-domain spectra were found in both Fig. 2A and Fig. 2B. The largest and second largest peaks were located at different frequencies for different TABLE 2 Partial correlation coefficients between grinding parameters
P F t C fso Pso
P
F
t
C
F8o
Pso
-1.0000 -0.1340 -0.1626 0.1059 -0.3920 -0.3045
- 1.0000 -0.8724 0.8028 -0.5259 -0.3034
- 1.0000 0.9785 -0.3862 -0.7167
- 1.0000 0.3323 0.7728
-1.0000 -0.1076
- 1.0000
P is power draw, F is feed rate, t is pulp temperature, C is pulp density, and F8o and P8o are 80% passing size for the mill feed and product respectively.
MONITORING GRINDING PARAMETERSBY SIGNAL MEASUREMENTS:INDUSTRIAL BALLMILL
1.4
7
Z27
(a) Acoustic signal 1.0 ~ 0.6
~ 0.2 ,~
I l,,,i
<
280 337
(b) Vibration signal
6
4
285
337400459
^ AA~7. 500
1216 tl
1000
5oo
Frequency (Hz) Fig. 2. Example original spectra of b o t h v i b r a t i o n a n d acoustic signals.
sensors. The signals picked up by both microphone and vibrometer built up a "stereograph" for grinding. The microphone picked up the signal emitted from the mill shell on the abrasion zone, and the vibrometer sensed the signal coming through three bearings, which was dispersed and disturbed by the bearing; therefore the microphone may receive more information than the accelerometer. Since mechanical system and grinding media generated a strong "finger print" spectrum for the grinding system, variations caused by varying operating parameters were buried in the main spectrum or presented as a modification to the main spectrum. Due to the wear of grinding balls, liners and bearings, it is difficult to know the absolute "finger-print" spectrum of the mill. Therefore variations of spectra are more important than the absolute one. To evaluate variations of spectra under different grinding conditions, the average spectrum of 40 observed spectra was adopted as a basis for comparison. Fig. 3 shows variations of five power spectra of vibration signal under eight different grinding conditions in a waterfall plot (Zeng and Forssberg, 1992 ). Although a wide range of frequencies between 50 and 1500 Hz are included, the significant variations of the spectra are situated at 50 and 450 Hz. Fig. 4 shows zoomed frequency bands between 500 and 1500 Hz. Some small variations are located on the frequency bands at 500-600 and 1200-1300 Hz. Fig. 5 shows variations of power spectra of the acoustic signal under eight
8
Y. ZENGAND E. FORSSBERG
8
[ Waterfall step : 11 vib08
L
~
6
vib07
e~ ~
vib06 4
- -
vib05 vib04 vib03
<
vib02
0
vibO1 560
1600
1500
Frequency (Hz) Fig. 3. Full spectra of vibration signals in waterfall plot.
0.4
Waterfall step : 0.05 I -
-
~
vib08
k,
vib07
-~ 0.3
vib06 vib05
0.2 ..
A_.__
vib04
*l,,m
vib03
0.1
.< ~,
~
vib02 vibO1
600
800
1600
1200
1400
Frequency (Hz) Fig. 4. Zoomed spectra on high frequency band of vibration signals.
'
MONITORINGGRINDINGPARAMETERSBYSIGNALMEASUREMENTS:INDUSTRIALBALLMILL 1.6 Waterfallstep=0.2 i
~
9
mic08 mic07
~- 1.2 ~~y-~-
mic06
~
0.8
mic05 mic04
~ - - ~
•- 0.4
mic03 mic02
<
~
0 5()0
micO1
1 oo
101)0
Frequency (Hz) Fig. 5. Fullspectraof acousticsignalsin waterfallplot. 0.8 Wateriallstep=O.l[' ~ t_
mic08
•
0.6
mic07
.
.
~
mic06 mic05
0.4
mic04 ~
~. 0.2
mic03 mic02
0
micO1
6bo
i
800
1600
1200
i
1400
F r e q u e n c y (Hz)
Fig. 6. Zoomedspectraon highfrequencybandof acousticsignals. different grinding conditions in a waterfall plot. The main variations of signal energies are within 50-500, 700-750, 1150-1200, 1350-1400 Hz. Fig. 6 shows a zoomed picture of the frequency band at 500-1500 Hz. Three frequencies
10
Y. ZENG AND E. FORSSBERG
of significant variations are at 725, 1172 and 1365 Hz. Comparing Fig. 5 to Fig. 3 and Fig. 6 to Fig. 4, the individual peak energy from the vibrometer was larger than that from the microphone. However, more peaks were found in the acoustic spectrum. The reason was that the signal caused by grinding was mixed with the signal from the bearings. The frequency band at 1200-1400 Hz mainly denoted the movement of mill charge (mixture of steel balls and iron ore pulp ). Since the microphone was situated at the abrasion zone, the acoustic pressure change can be detected directly by a sensitive microphone; however, it will be dispersed on travelling from the mill body to the vibration sensor through bearings and pinion axis. Therefore, the high frequency vibration signal from grinding cannot travel efficiently through the gears and bearings.
Identification of vibration and acoustic signals Principal component analysis is a useful technique for reducing variables in a data set by finding linear combinations of these variables that explain most of the variability. The purpose of this is to express p variables in X1, )(2, ..., Xp with a group of new indices Z1, Z2, ..., Zk that are uncorrelated, and where k << p. Moreover, the indices are ordered so that ZI displays the largest amount of variation of the data set, Z2 displays second largest amount of variation, and so on. Here Zi is called the ith principal component (a latent variable ). Since energy levels of peaks on spectra are highly correlated, principal component analysis is suitable for spectrum identification. Before applying principal component analysis, the original spectral data were centred, i.e., each observation of a certain variable must subtract the mean of all observations of that variable. Principal component analysis was performed with all original variables. The critical values at 0.01 and 0.02 were set to remove less significant variables for the vibration and acoustic signals respectively. After principal component analysis, 73 and 116 original frequency elements were relevant to vibration and acoustic signals respectively. Fig. 7 shows weights of significant frequency elements for the first four principal components of the vibration signal. The first, second, third and fourth principal components account for 85.2, 8.3, 2.8 and 1.2% of the total variances of the original frequency elements respectively. Fig. 7A shows that the first principal component depended mainly on 112 and 117 Hz and the second on 283 and 117 Hz. Fig. 7B shows that the third principal component was related to 171 and the fourth to 288, 171 and 112 Hz. Fig. 8 shows weights of the significant frequency elements for the first four principal components of the acoustic signal. The first, second, third and fourth principal components account for 50.2, 24.5, 7.1 and 4.2% of the total variances of the original variables respectively. Fig. 8A shows that the first and
MONITORING GRINDING PARAMETERS BY SIGNAL MEASUREMENTS: INDUSTRIAL BALL MILL
0.5
W = Weight of component
177
(A)
1P2
0
q~
W2
°°~ °
288
-0.5 2~3
-1.0 -0.2
I
0
o16
0.4
0'.2
0.8
W1 0.4
(B) o
0
W4
o.~°:
°
4o~
4~9
-0.4 288 o
I
-0.2
-0.4
0~2
0
0'.4
0'.6
0.8
W3 Fig. 7. Weight of first four principal components of vibration signals.
W = Weight of component
7~
0.4 (A) 6s
W2
ooO O~oO117
o
112 o ~225
-0.4
229 n
-0.2
6
On
0.2
0.4
016
0.8
W1 0.8
(B) 117
0.4
172
171
W4
1172 o
0
o~j,:~o 8 v
o ° 1362 ~
o o
o
1367
225 o
229
-0.4
-0.2
6
012
0'.4
0..6
W3 Fig. 8. Weight of first four principal components of acoustic signals.
0.8
11
Y. ZENG AND E. FORSSBERG
12
second principal components were mainly contributed by 229 and 73 Hz. Fig. 8B shows that the third principal component was contributed by 1367, and the fourth 112 and 117 Hz.
Model developmentfor grinding parameters After principal component analysis, four latent variables can describe most of the variations of the original spectra. The scores can be calculated if the weights of principal components and measured power spectra from grinding are known. The scores of vibration and acoustic signals were denoted by Svi and Smi respectively, where i denotes the score on the ith principal component. By multiple regression analysis, grinding parameters can be related with the scores, that yields: P = 1107.6 + 7.72Svl - 15.23Sv2 + 24.4Srnl -- 29.2Sm2 - 63.5Sm6
( 1)
F = 191.6 -
(2)
58.73Sv3
-
109.4Sm~ + 90.1Sm2 -- 71.4Sm3 + 107.3Sm5
C = 6 7 . 8 - 3.92Sv~ + 6.44Sv4 + 3.23S~1 - 6.65Sm4 - 17.6Sm6
(3)
P8o =0.271 - 0.0443Sv3 -0.140Sml + 0.128Sm2 - 0.071Sin3 + 0.218S~6
(4) where P denotes power draw (kW), F feed rate (t/h), C pulp density (% solids by weight) and P8o mill product size (mm). Equations (1-4) account for 72.5, 93.6, 84.9 and 94.3% of the corresponding observations. Eq. ( 1 ) shows that power draw is mainly related to the first two principal components of vibration and acoustic spectra. Fig. 9 shows observed and predicted power draw. The prediction clearly shows the variation tendency of power draw. The prediction is better in a high feed rate than in a low one. At equilibrium state in grinding, the higher the feed rate, the more material in the mill, and the smaller the collision probability between grinding media. The variation of power draw with the higher feed rate was smaller than with the lower one. The prediction can be improved by implementing a more accurate power meter for calibration purposes. Eq. (2) shows that the feed rate to the ball mill is mainly related to acoustic spectra. The feed rate was affected by the third component of vibration and acoustic spectra. Fig. 10 shows observed and predicted feed rate to the ball mill. Two levels of feed rates are clearly separated; however, the prediction error is larger than the natural variation of the feed rate. The material fed into the ball mill varied since the weighed material was first fed to the magnetic separator, and not to the mill directly. Since the properties of the feed material vary with time, the output from the magnetic separator varies correspondingly. The real feed rate to the ball mill has to be measured to improve prediction accuracy.
MONITORING GRINDING PARAMETERSBY SIGNALMEASUREMENTS:INDUSTRIALBALLMILL
13
1160 * --
Observation 1 Prediction
~: 1140
~ 1120
~ 1100
1080 1'0
20
3o
40
Time index Fig. 9. Observed and predicted power draw by the ball mill.
260 * --
Observation ] Prediction
,t~ 220
,40
"~ 180
0
1'0
20
3¢)
40
Time index Fig. 10. Observed and predicted mill feed rate to the ball mill.
The ground product from mill discharge was randomly sampled during the valid testing period. The sampled material was used to determine both pulp density and ground product size. Eq. (3) shows that the pulp density was
14
Y.ZENGANDE. FORSSBERG
~.~ 72 l
70 "~ 68 .~,,9
~ 66 e,~ 64 62 0
1'0
--•-T 20
I
*
Observation Prediction 3~)
4O
Time index Fig. 11. Observed and predicted pulp density inside the ball mill. ~-~ 0.36 [*Observation] I
@. 0.32 o~ ¢d
0.28
"~ 0.24
~ 0.2o 0
1'0
20
3~)
40
Time index Fig. 12. Observed and predicted ground product size of mill discharge. affected by both the first and fourth components from vibration and acoustic signals. Fig. 11 denotes observed and predicted pulp densities o f mill discharge. The prediction follows the variation tendency of pulp density. The al-.,olute error of prediction is within 2%.
MONITORING GRINDING PARAMETERS BY SIGNAL MEASUREMENTS: INDUSTRIAL BALL MILL
15
Eq. (4) shows that the ground product size is mainly related to acoustic spectra. The product size was affected by the third component of both vibration and acoustic spectra. Fig. 12 shows the observed and predicted 80% passing size of ground product. The prediction can satisfy most of the requirements in practice. The key grinding parameters in ball mill grinding can be predicted by a combination of principal components derived from vibration and acoustic spectra. Particularly, feed rate and product size can be satisfactorily predicted based on the acoustic signal. CONCLUSION
Vibration and acoustic signal measurements were made on an industrial scale primary ball mill at LKAB, Malmberget, Sweden. Three variable operating parameters were considered: feed rate, feed size and pulp density. The measured response parameters were product size, power draw and pulp temperature. Partial correlation analysis shows that pulp density and pulp temperature are highly correlated. Power draw, feed rate, pulp density and ground product size were four key grinding parameters. The mechanical vibration from the grinding system was picked up by an accelerometer, and the acoustic pressure changes at the centre of the abrasion zone outside the mill shell were sensed by an acoustic microphone; thus a "stereograph" can be constructed for signals from the ball mill. The original signals were stored on a DAT deck during the valid testing periods. After signal processing and spectral identification, relationships between key grinding parameters and signal characteristics were established by principal component analysis and multiple regression. The feed rate and ground product size were mainly related to the acoustic signal. Power draw and pulp density can be predicted by a combination of vibration and acoustic signals. All the key grinding parameters can be monitored by measuring signals emitted from an industrial scale ball mill. Monitoring key grinding parameters by signal technique can provide more information on grinding to the control system, to ensure it works more efficiently. ACKNOWLEDGEMENT
This work was partially sponsored by the Norrbotten's Research Board and LKAB, Sweden. The authors thank Kent Tano and Jan-Christer G/irde of LKAB, Malmberget for their assistance during the industrial tests.
16
Y. ZENGANDE. FORSSBERG
REFERENCES Anon., 1990. 386 - - MatLab User's Guide. The Math Works Inc., South Natick, MA, 15 October. DeFatta, D.J., Lucas, J.G. and Hodgkiss, W.S., 1988. Digital Signal Processing: A System Design Approach. Wiley, Toronto, Vol. 29, pp. 394-401. Little, J.N. and Shure, L., 1988. Power spectral estimation. In: Signal Processing Toolbox. The Math-Works Inc., August, 1: 37-41; 2: 43; 2: 73. Thomas, D.W., 1978. Vehicle sounds and recognition. In: B.G, Batchelor (Editor), Pattern Recognition: Idea in Practice. Plenum Press, New York, pp. 333-339. Welch, P.D., 1967. The use of fast Fourier transform for the estimation of power spectra. IEEE Trans. Audio Electroacoust., June, AU-15: 70-73. Wold, S., Esbensen K. and Geladi P., 1987. Principal component analysis. Chemometrics Intell. Lab. Syst., 2: 37-52. Zeng, Y. and Forssberg, E., 1992. Effects of operating parameters on vibration signal under laboratory scale ball grinding conditions. Int. J. Miner. Process., 35: 273-290. Zeng, ¥. and Forssberg, E., 1993. Application of digital signal processing and multivariate data analysis to vibration signals from ball-mill grinding. Trans. Inst. Min. Metall. (Sect. C: Miner. Process. Extr. Metall. ), 102 (January-April): 39-43.